Hi I have been going though the Vitis ai where you feed in your frozen (.pb) model and it gives you a corresponding bit stream using the DPUs. I find many of the layers, like say LSTM are not available as of now. Suppose my custom model contains many supported layers and and couple of un-supported ones, is there a way to use these DPUs like an IP for the supported layers? If so, how do I do it? Thanks in advance Ashwin Krishnan
The NN i am trying to implement contains upsampling layer which is currently not supported in U280 boards. I am considering to split the network layers where pre upsampling layers will be implemented on vitis-ai DPUs and remaining layers as hls_ip. Is such a configuration possible?
Does vitis-ai exports the bitstream (or some IR) that can be instantiated as a kernel in vitis flow.